From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from mailman by lists.gnu.org with tmda-scanned (Exim 4.43) id 1KQ497-0003Hc-2h for qemu-devel@nongnu.org; Mon, 04 Aug 2008 13:47:01 -0400 Received: from exim by lists.gnu.org with spam-scanned (Exim 4.43) id 1KQ495-0003HG-IK for qemu-devel@nongnu.org; Mon, 04 Aug 2008 13:47:00 -0400 Received: from [199.232.76.173] (port=39126 helo=monty-python.gnu.org) by lists.gnu.org with esmtp (Exim 4.43) id 1KQ495-0003HD-Cx for qemu-devel@nongnu.org; Mon, 04 Aug 2008 13:46:59 -0400 Received: from an-out-0708.google.com ([209.85.132.242]:18871) by monty-python.gnu.org with esmtp (Exim 4.60) (envelope-from ) id 1KQ495-0000y3-D5 for qemu-devel@nongnu.org; Mon, 04 Aug 2008 13:46:59 -0400 Received: by an-out-0708.google.com with SMTP id d18so644938and.130 for ; Mon, 04 Aug 2008 10:46:58 -0700 (PDT) Message-ID: <48974066.1090803@codemonkey.ws> Date: Mon, 04 Aug 2008 12:46:14 -0500 From: Anthony Liguori MIME-Version: 1.0 Subject: Re: [Qemu-devel] [PATCH 5/7] xen: add block device backend driver. References: <1217865045-10722-1-git-send-email-kraxel@redhat.com> <1217865045-10722-6-git-send-email-kraxel@redhat.com> In-Reply-To: Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 7bit Reply-To: qemu-devel@nongnu.org List-Id: qemu-devel.nongnu.org List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: qemu-devel@nongnu.org Cc: xen-devel@lists.xensource.com, Gerd Hoffmann Blue Swirl wrote: > On 8/4/08, Gerd Hoffmann wrote: > >> +/* i386 protocol version */ >> +#pragma pack(push, 4) >> > > What's wrong with __attribute__(__aligned__)? > > >> + * FIXME: the code is designed to handle multiple outstanding >> + * requests (using aio or using threads), which isn't used right >> + * now due to limitations of the qemu block driver interface. >> > > Yes, what happened to vectored DMA patches? > Believe it or not, they were hurting performance in KVM. Passing a vector to the block layer will force the IO operation to be split up into multiple requests that are all executed in serial (since QEMU only supports 1 outstanding request at a time). Copying the vectored IO to a linear buffer increased performance significantly because it avoided this serialization. The block layer needs some refactoring to support proper vectored AIO operations before a zero-copy API will make sense from a performance perspective. Regards, Anthony Liguori